Data Throughput for Efficient Photonic Neural Network Accelerators

Machine Learning has become a dominant technology, spurring the invention of photonic systems to implement neural network tasks. These photonic systems offer high throughput operations with low power but require large bandwidths of data to operate with maximum efficiency. Here we provide an analysis of the required data bandwidths, reaching nearly 1 Tbps for a single chip and necessitating the use of high bandwidth memory.